Hepatocellular carcinoma (HCC) is the most common primary liver cancer. It arises predominantly in patients
with chronic liver disease or cirrhosis, a population that undergoes routine surveillance imaging. Critically,
HCC is one of the few cancers where imaging alone can establish a diagnosis; no biopsy is required.
What is LI-RADS?
LI-RADS (Liver Imaging Reporting and Data System) standardizes the description and categorization of liver
observations on contrast-enhanced CT and MRI in patients at high risk for HCC. It assigns observations to
categories based on imaging features.
The framework uses temporal enhancement across multiple contrast phases — non-contrast, arterial, portal
venous, and delayed — to characterize how a lesion changes relative to the surrounding liver.
The Annotated Multi-Phase Liver Imaging for Artificial Intelligence (AMPLIFAI) Challenge was organized as a
MICCAI 2026 benchmark for clinically meaningful AI in liver lesion assessment on multi-phase CT. AMPLIFAI
translates the LI-RADS clinical framework into a reproducible lesion-level classification benchmark.
Benchmark task
Participants predicted the LI-RADS category of one pre-defined target lesion per case from multi-phase CT.
Final challenge ranking was based solely on LI-RADS classification performance.
Dataset & challenge design
Training data were harmonized and preprocessed from four source datasets and included multi-phase CT,
lesion-level LI-RADS labels, radiologist-provided lesion segmentations, and voxel-level annotations of key
imaging features including non-rim APHE, non-peripheral washout, and enhancing capsule. These annotations
supported both end-to-end and feature-guided approaches but were not separate challenge endpoints.
LI-RADS defines specific enhancement features that, when present, raise concern for HCC. Four major features are
evaluated:
Arterial phase hyperenhancement (APHE) - the lesion enhances brighter than the surrounding
liver in the early arterial phase, reflecting the abnormal arterial supply that characterizes HCC. Only
non-rim APHE counts - enhancement limited to the outer edge of the lesion (rim pattern) suggests a
different diagnosis.
Non-peripheral washout - in the portal venous or delayed phase, the lesion becomes relatively
darker than the surrounding liver. This reflects contrast washing out of the tumor faster than the background
liver parenchyma.
Enhancing capsule - a thin bright rim persisting at the lesion edge in the portal venous or
delayed phase. This reflects contrast being retained in the fibrous tissue that surrounds and encases the
tumor.
Threshold growth - a significant increase in lesion size over a short interval (≥50% diameter
growth in ≤6 months), reflecting rapid tumor proliferation (not present as a label in the challenge dataset).
Confirming HCC from imaging requires placing a lesion into the LR-3, LR-4, or LR-5 range - which reflects
increasing probability of HCC from intermediate to definite. The algorithm for reaching those categories is built
on the features above.
How are LR-3 through LR-5 determined?
The primary gating feature is non-rim APHE. Lesions without non-rim APHE can reach a maximum of LR-4,
regardless of size or other features. Lesions with non-rim APHE are then categorized by a combination of
size and how many additional major features are present (washout, capsule, or threshold growth) - more features and
larger size push the category higher, from LR-3 up to LR-5.
